Jørgensen Mathias S, Groves Michael N, Hammer Bjørk
Interdisciplinary Nanoscience Center (iNANO) and Department of Physics and Astronomy, Aarhus University , Aarhus DK-8000, Denmark.
J Chem Theory Comput. 2017 Mar 14;13(3):1486-1493. doi: 10.1021/acs.jctc.6b01119. Epub 2017 Feb 22.
Predicting structures at the atomic scale is of great importance for understanding the properties of materials. Such predictions are infeasible without efficient global optimization techniques. Many current techniques produce a large amount of idle intermediate data before converging to the global minimum. If this information could be analyzed during optimization, many new possibilities emerge for more rational search algorithms. We combine an evolutionary algorithm (EA) and clustering, a machine-learning technique, to produce a rational algorithm for global structure optimization. Clustering the configuration space of intermediate structures into regions of geometrically similar structures enables the EA to suppress certain regions and favor others. For two test systems, an organic molecule and an oxide surface, the global minimum search proves significantly faster when favoring stable structures in unexplored regions. This clustering-enhanced EA is a step toward adaptive global optimization techniques that can act upon information in accumulated data.
在原子尺度上预测结构对于理解材料的性质非常重要。没有高效的全局优化技术，这样的预测是不可行的。许多当前技术在收敛到全局最小值之前会产生大量闲置的中间数据。如果在优化过程中能够分析这些信息，那么对于更合理的搜索算法就会出现许多新的可能性。我们将进化算法（EA）和聚类（一种机器学习技术）相结合，以产生一种用于全局结构优化的合理算法。将中间结构的构型空间聚类成几何结构相似的区域，使得进化算法能够抑制某些区域并偏向其他区域。对于两个测试系统，一个有机分子和一个氧化物表面，当偏向未探索区域中的稳定结构时，全局最小值搜索明显更快。这种聚类增强的进化算法是朝着能够根据累积数据中的信息采取行动的自适应全局优化技术迈出的一步。
